【摘要】 车辆分类是智能交通系统的关键技术之一，是道路交通监控系统的一个重要研究领域。由于声学传感器具有效率高、成本低、可昼夜工作、隐蔽性强等优势，因此基于车辆声音特征的车辆分类引起了研究人员的广泛关注。然而，现有研究中的车辆声音信号仅包含单一车辆，对于混合的双车辆声音信号的分类缺乏讨论。为此，设计一种网络模型对单车辆和双车辆共12种类别的噪声信号进行分类。针对声音频谱特征的固定分辨率并非最优的问题，基于网络训练得出的注意力得分和时间转换矩阵，控制噪声频谱时间大小，设计频谱时间分辨率优化模型。分类网络依据卷积递归神经网络(CRNN)架构，卷积网络部分(多尺度信号分析模块)参考高效空间金字塔模块对特征进行双分支融合处理，由于循环神经网络(RNN)等不利于并行化，运算速度慢，因此将因果时间卷积网络(TCN)转换为非因果循环TCN。在自制数据集中进行实验，结果表明，该模型的平均精度均值(mAP)达到0.98,远高于相当参数量下的CRNN网络，与MobileNetV3性能相当，但是相比MobileNetV3参数量减少了1.7×10~6。分析结果表明，所提模型适用于长时间声音信号处理任务，能提取深层次的特征。更多还原

【Abstract】 Vehicle classification is a key technology in intelligent transportation systems and a vital research area in road traffic monitoring systems. Owing to the advantages of acoustic sensors, such as high efficiency, low cost, round-the-clock operations, and strong concealment, vehicle classification based on vehicle sound characteristics has been extensively researched. However, existing vehicle sound signals only contain a single vehicle, with limited discussion on classifying mixed two-vehicle sound signals. To address this research gap, a network model is developed to classify noise signals from single and double vehicles. To address the issue of suboptimal fixed resolution in sound spectral features, a spectral time-resolution optimization model is designed using the attention score and frame warpage matrix obtained from network training. The classification network is based on a Convolutional Recurrent Neural Network(CRNN) architecture, with the convolutional component(multiscale signal reconstruction module) utilizing an efficient spatial pyramid for double-branch fusion. Since Recurrent Neural Network(RNN) and other cyclic networks are unsuitable for parallelization and have low operation speeds, the causal Time Convolutional Neural Network(TCN) is converted to a non-causal cyclic TCN. The mean Average Precision(mAP)of the model on the self-made dataset reaches 0.98, significantly outperforming the CRNN network with a comparable parameter count. Its performance is comparable to MobileNetV3 but with 1.7×10~6 fewer parameters. Experimental results indicate that the designed model is effective for processing long-term sound signals and extracting deep features.更多还原